Machine‐learning based spatiotemporal prediction of soil moisture in a grassland hillslope

Abstract Soil moisture (SM) plays a significant role in the earth's water balance and in optimizing land management practices. However, SM at the field scale is difficult to map from available point measurements due to the inherent heterogeneity of soil and terrain properties and temporal dynamics of weather conditions. In this study, we explored the potential of four machine learning (ML) methods (random forest, gradient boosted regression trees, support vector regression, and neural networks) to predict SM in a grassland hillslope in space and time using auxiliary variables on soil and terrain properties and weather conditions. For training and testing the ML models, we used SM point measurements obtained by a sensor network. Performance metrics varied between the ML methods and the training‐test data split ( R 2 = 0.48–0.69, root‐mean‐square error [RMSE] = 0.06–0.10). Random forests and gradient‐boosted regression trees turned out to be promising and easy to parametrize as first choices to explore the potential of ML techniques. The day of the year emerged as an important feature to predict SM across models and can thus serve as a proxy for seasonal hydroclimatic variability. To enable the transfer of the application to other contexts or sites, we provide the modeling workflow as an open‐source computational Python module.